Agentic AI: Practical Adoption, Governance, and Measurable Impact
Agentic AI is being marketed as the next major leap beyond chatbots, yet many organizations remain stuck in pilots with limited measurable outcomes.
In a recent podcast-style conversation with LinkedIn’s Ray Villalobos, Andreas Welsch—an AI leadership expert focused on AI strategy, governance, adoption, and workforce transformation—described what is working in practice today, what is mostly hype, and how leaders can move from experimentation to business impact without creating unnecessary risk.
His perspective also addresses a core executive tension: employees increasingly use AI tools with or without approval, while leadership teams must protect data, manage ROI expectations, and build a culture that enables responsible adoption.
Executive Summary
- “Agent washing” blurs the line between workflow automation and true agents.
- True agents can adapt, reason, and act—beyond linear “then-do” workflows.
- Adoption is uneven: AI-mature firms move faster; others remain cautious.
- Most AI initiatives fail without data readiness, culture, and realistic milestones.
- Leaders should enable experimentation with clear guardrails, not fear-based mandates.
Key Takeaways
- Andreas Welsch distinguishes workflows (“first do this, then that”) from agents that can pursue a goal and adapt along the way.
- Vendor hype is a predictable market force, but enterprise adoption typically moves slower due to people, processes, and governance.
- AI-first mandates can create resistance when employees lack baseline AI literacy and confidence.
- AI projects are cyclical and research-like; leaders should expect iteration, setbacks, and controlled “stop” decisions.
- “Throwing spaghetti at the wall” with many AI subscriptions rarely outperforms a few well-chosen bets tied to business KPIs.
- Multi-agent systems (agent “swarms”) are emerging, supported by interoperability approaches like tool-calling and agent-to-agent communication.
- Quality and accountability still matter; leaders should actively prevent “AI slop” from becoming acceptable output.
What is Agentic AI?
Agentic AI refers to AI systems that can accept a goal, break it into subtasks, work through those tasks, and return recommendations—and in some cases take action. Andreas Welsch differentiates this from classic workflow automation, where steps are predefined (for example: create a document, generate an outline, save it, send it). A “true agent,” in his framing, can be brought into collaboration tools like Slack or Teams, pull information from connected sources such as the internet, summarize it, and adapt when the original plan fails—functioning more like a digital employee operating under human direction.
Why this conversation matters
The conversation reflects a reality facing CIOs, CTOs, CHROs, and business leaders: the market promises rapid transformation, while internal constraints—risk, data protection, process maturity, and AI literacy—slow the path from pilots to outcomes.
Andreas Welsch’s focus is executive-relevant because it connects agentic capabilities to practical adoption patterns, governance choices, and workforce transformation. It also names the cultural failure modes (fear-based mandates, unrealistic expectations, and low accountability for output quality) that can derail even well-funded initiatives.
Agent washing: why “agents” often aren’t agents
Welsch describes “agent washing” as a market dynamic where vendors label products as “agents” even when the functionality is closer to rebranded automation. In his view, this follows a familiar technology hype pattern: vendors want attention, startups want funding, and product marketing highlights the most ambitious promise.
The executive risk is not merely semantic. When leaders assume “agent” implies adaptive reasoning and autonomous task execution, they may overestimate value, underestimate governance needs, and set expectations the organization cannot meet.
Key Insight: Andreas Welsch explains that “agent washing” inflates expectations by branding workflows as agents, pushing leaders to buy capabilities they do not actually receive. Clear definitions protect budgets, timelines, and trust—especially when employees already feel pressure to “use AI” without clarity on what counts as success.
Workflow automation vs. true agents: the operational distinction
Welsch separates two categories often conflated under “agentic AI.” The first is workflow automation: pre-sequenced steps with an AI task embedded (for example, creating an outline or drafting text), followed by the next deterministic step. This model is “first do this, then do that.”
The second is a true agent that can be invoked in context, such as a “research agent” in Slack or Teams that can find financial information about companies, summarize it, and return insights relevant to the conversation. In this model, the agent can adapt—if one source fails, it can pursue alternatives—because it is working toward a goal rather than executing a strict script.
Key Insight: In Andreas Welsch’s definition, a workflow follows predetermined steps, while a true agent is goal-driven and adaptive. That difference affects governance, ROI measurement, and operating model decisions because agents may need broader tool access, clearer delegation instructions, and stronger accountability checks.
Where most companies actually stand on Agentic AI adoption
Welsch observes a wide adoption spectrum. Large multinational organizations that have used machine learning, predictive analytics, and statistics for years have often built the organizational “muscle” to evaluate AI opportunities, manage data dependencies, and decide which initiatives are worth investment.
Other companies are still taking a “wait and see” approach—watching peers to avoid expensive lessons and then investing in a limited set of strategic use cases. Welsch links slower adoption to typical enterprise realities: existing processes, governance requirements, and the need to bring people along, not just deploy tools.
He also notes that employees frequently use AI tools regardless of official approval. That makes the gap between “allowed tools” and “market availability” a governance issue as much as a technology issue.
Multi-agent systems: what executives should watch next
Welsch expects more “agent swarms” or multi-agent systems, where multiple specialized agents collaborate on a single outcome. He gives a marketing example: one agent researches an ideal customer profile, another drafts messaging, another generates creative and ad copy, and a final agent checks alignment with brand and strategy guidelines.
He also points to emerging technical enablers: agent-to-agent protocol and model context protocol as examples of how agents may call tools and communicate more effectively. The practical implication is that interoperability will increasingly shape which platforms scale inside the enterprise.
Key Insight: Andreas Welsch anticipates that organizations will move beyond single-purpose tools toward multi-agent collaboration. The winners will be the teams that design agents like teams—clear roles, shared context, and explicit quality checks—rather than expecting “autonomy” to replace management.
Staying safe without losing competitiveness: a lightweight governance approach
Welsch describes a common executive dilemma: leaders want to be forward-leaning, but they must avoid exposing sensitive data or creating compliance and security risks. He cautions against two extremes, especially for mid-sized organizations (for example, 200–300 employees): overengineering governance with slow review boards, or allowing a free-for-all where anyone can adopt any tool.
His practical recommendation is to create a simple framework with clear “don’t dos” (high-risk use cases that are never allowed) and categories of acceptable tools—such as enabling AI features already included in licensed products, and separately evaluating tools IT and developers need to work efficiently.
He also recommends a lightweight intake process: a short survey to capture the use case, department, expected KPI improvement (productivity or another metric), and cost. After a quarter or two, leaders should revisit whether expected value was realized and decide whether to continue, switch, or stop.
Culture of trust vs. culture of fear: why “AI-first or else” backfires
Welsch challenges fear-based messaging. He notes that some CEOs have publicly pushed “AI-first…or else” narratives, which can trigger resistance. In contrast, he sees more effective approaches that invite experimentation, provide approved tools and resources, and establish communities of practice where employees share what works and what does not.
He stresses that many organizations are not yet “AI-ready.” Employees often still treat tools as a question-and-answer machine—“more than Google”—without understanding limitations, opportunities, and responsible usage. In his view, basic AI literacy is a prerequisite for meaningful adoption, particularly before leaders expect workflow redesign or productivity transformation.
Key Insight: Andreas Welsch argues that AI adoption is a culture and enablement challenge, not just a technology rollout. Trust-based programs—approved tools, learning resources, and communities of practice—help employees build confidence and reduce shadow AI usage that increases risk.
Why AI projects fail: linear project thinking meets cyclical reality
Welsch points to repeated industry signals that a large share of AI initiatives do not deliver positive outcomes. He highlights a dynamic leaders often underestimate: AI projects are not linear “start-to-finish” efforts. They are cyclical and iterative, closer to research projects than classic implementation projects.
In practice, teams move through repeated cycles of idea generation, data acquisition, hypothesis testing, and validation, with only some projects ever reaching production. Welsch recommends setting review points where technical teams and business sponsors assess progress every few weeks, identify blockers, request business input, and decide whether to proceed, extend, pause, or stop.
He also emphasizes that pulling the plug should be treated as a learning outcome—capturing what was tried, what failed, and what should change next time—rather than a blame event.
Small and localized models: emerging options leaders are discussing
In a Switzerland-based discussion, Welsch recounts questions from senior leaders about bias and “world views” embedded in different model families. The group explored whether different models might be needed for different purposes and contexts.
He points to a trend toward smaller language models trained on curated subsets of data to meet more specific needs. He also notes announcements about localized or regional large language model initiatives, including Switzerland’s announcement of an open, transparent model developed by universities there, and similar efforts referenced in Australia.
Preventing “AI slop”: quality and accountability still define leadership
Welsch describes an emerging management challenge: AI-enabled output can be “not bad,” but still fall below professional standards. He references the idea of “AI slop” (or “AI workslop”)—content that appears unreviewed, minimally edited, and overly dependent on a tool’s first draft.
His position is direct: AI availability does not remove accountability for quality. Leaders must set expectations for how AI is used, how outputs are reviewed and refined, and how teams remain responsible for final deliverables.
Leadership Implications
- Define “agent” inside the enterprise. Use Welsch’s workflow-versus-agent distinction to prevent agent washing from driving spend and expectations.
- Adopt lightweight AI governance. Establish clear “never do” rules, categorize low-risk activations vs. higher-risk tools, and avoid slow review boards.
- Measure value with simple KPIs. Use a short intake survey for tool requests and review outcomes after a quarter or two.
- Build AI literacy before mandates. Replace “AI-first or else” pressure with enablement, resources, and communities of practice.
- Run AI projects as iterative programs. Set recurring review checkpoints, expect nonlinearity, and treat stop decisions as learning, not failure.
Conclusion
Agentic AI is progressing quickly in product marketing and tooling, but Andreas Welsch’s perspective highlights the practical leadership work required to translate those capabilities into measurable outcomes. The path forward depends on clear definitions, lightweight governance, AI literacy, iterative project management, and strong accountability that prevents AI workslop from becoming a new standard.
For executives, the opportunity is not merely adopting Agentic AI. It is building an operating model where agents, workflows, and people collaborate safely—while delivering quality that strengthens trust with customers and within the workforce.
FAQ
1) What is the difference between an AI workflow and Agentic AI?
A workflow follows predefined, linear steps (do A, then B, then C), even if one step uses generative AI. Agentic AI, as Andreas Welsch describes it, is goal-driven: it can break work into subtasks, adapt, and return recommendations or actions.
This distinction matters because governance, tool access, and accountability requirements increase as systems become more agent-like.
2) What does “agent washing” mean in enterprise AI?
Agent washing is when vendors market products as “agents” even when the underlying capability is primarily workflow automation or rebranded functionality. Andreas Welsch notes this often happens during tech hype cycles and can inflate expectations inside organizations.
Leaders can reduce confusion by defining what counts as an agent and validating capabilities against real use cases.
3) Why are so many AI pilots failing to produce measurable results?
Many AI initiatives fail because leaders treat them like linear projects, underestimate data and process dependencies, and set unrealistic expectations from vendor hype. Andreas Welsch emphasizes that AI projects are cyclical, research-like, and require iterative checkpoints and learning—sometimes including stopping.
Success improves when business sponsors stay engaged and outcomes are evaluated against clear KPIs.
4) How can a company stay competitive on AI while managing data risk?
Organizations can stay competitive by using a simple governance framework that defines high-risk “don’t dos,” enables low-risk AI features in existing licensed tools, and evaluates new tools with lightweight review. Andreas Welsch cautions against both slow review boards and uncontrolled adoption.
A short intake survey and quarterly value review helps balance speed and safety.
5) What is a multi-agent system, and why does it matter?
A multi-agent system (or “agent swarm”) uses multiple specialized agents that collaborate on an outcome, such as research, messaging, creative generation, and brand compliance checks. Andreas Welsch expects these systems to become more common as interoperability improves through tool-calling and agent-to-agent communication.
Executives should plan for new workflow design and stronger quality controls as agent collaboration increases.
6) Why can “AI-first or else” messaging create resistance?
Fear-based AI mandates can backfire because many employees are not yet AI-ready and may not understand limitations, best practices, or what “good” looks like. Andreas Welsch argues that trust-based enablement—tools, resources, and communities of practice—drives adoption more reliably than pressure.
This approach also reduces the likelihood of shadow AI usage that increases risk.
7) What is “AI slop” (AI workslop) and how should leaders respond?
AI slop refers to content that looks like an unedited first draft from a tool—acceptable at a glance but below professional standards. Andreas Welsch emphasizes that AI does not remove accountability for quality, so leaders must set expectations for review, editing, and ownership of outputs.
Clear quality bars protect credibility with customers and internal stakeholders.
8) How should leaders decide which AI tools to approve?
Andreas Welsch recommends a lightweight intake process: capture the use case, department, expected KPI improvement (productivity or other), and cost. Then reassess after a quarter or two to compare expected versus realized value and decide whether to continue, change tools, or stop.
This method discourages “spaghetti on the wall” purchasing and improves ROI accountability.
9) Do companies need different AI models for different regions or values?
In executive discussions Welsch describes, leaders raised concerns about bias and differing “world views” embedded in models from different providers. He points to trends toward smaller, curated models and localized or regional LLM initiatives, including an open, transparent Swiss university-developed model.
Model selection may increasingly consider context, governance, and stakeholder expectations.
10) What capabilities should executives look for in “true” AI agents?
Welsch’s “true agent” characteristics include goal acceptance, task decomposition, ability to retrieve information from connected tools or the internet, summarization, and adaptability when initial paths fail. These agents can be used inside collaboration tools like Slack or Teams and act on behalf of users.
Leaders should test these behaviors in controlled environments before scaling access and autonomy.